ICLR 2026 Papers — Page 21
International Conference on Learning Representations · 5356 papers
Gradient-Direction-Aware Density Control for 3D Gaussian Splatting
Zheng Zhou (Shanghai University of Engineering Science), Hongjian Zhan (East China Normal University)
GenerationComputational EfficiencyGaussian SplattingImage
🎯 What it does: Propose a dynamic density control framework GDAGS based on Gradient Direction Consistency (GCR) to improve splitting and cloning operations in 3D Gaussian Splatting, thereby enhancing image quality for novel view synthesis and reducing the number of Gaussians.
Gradient-Normalized Smoothness for Optimization with Approximate Hessians
Andrei Semenov (EPFL), Nikita Doikov (Cornell University)
OptimizationBenchmark
🎯 What it does: This paper proposes a new optimization framework called Gradient-Normalized Smoothness, and based on this concept, designs a gradient regularized Newton algorithm using approximate Hessian, which can achieve global convergence on both convex and non-convex problems;
Gradient-Sign Masking for Task Vector Transport Across Pre-Trained Models
Filippo Rinaldi (University of Modena and Reggio Emilia), Simone Calderara (Vector Institute)
Knowledge DistillationRepresentation LearningTransformerImageText
🎯 What it does: This paper proposes a method called GradFix, which transfers task vectors from an old pre-trained model to a new model through gradient sign masking; this method only requires a small number of labeled samples to complete the transfer without additional fine-tuning.
GradPCA: Leveraging NTK Alignment for Reliable Out-of-Distribution Detection
Mariia Seleznova (Ludwig-Maximilians-Universitat Munchen), Gitta Kutyniok (Ludwig-Maximilians-Universitat Munchen)
ClassificationAnomaly DetectionImage
🎯 What it does: Proposed a gradient-based OOD detection method called GradPCA leveraging the low-rank structure of gradients.
GradPruner: Gradient-guided Layer Pruning Enabling Efficient Fine-Tuning and Inference for LLMs
Wei Huang (Ant Group), Yinggui Wang (Ant Group)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical Data
🎯 What it does: Propose GradPruner, an efficient LLM fine-tuning and inference method that evaluates parameter importance using early fine-tuning gradients and performs hierarchical pruning and merging.
GRAM-DTI: Adaptive Multimodal Representation Learning for Drug–Target Interaction Prediction
Feng Jiang (University of Texas at Arlington), Rui Liao (Johnson & Johnson Innovative Medicine)
Representation LearningDrug DiscoveryTransformerSupervised Fine-TuningContrastive LearningTextMultimodalityBiomedical DataBenchmark
🎯 What it does: Proposed the GRAM-DTI framework, which learns unified drug-target representations through multi-modal pre-training for DTI prediction.
GranViT: A Fine-Grained Vision Model For Autoregressive Multimodal Large Language Models
Guanghao Zheng (Shanghai Jiao Tong University), Qi Tian (Huawei Inc)
RecognitionSegmentationKnowledge DistillationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Propose the GranViT model, combining fine-grained visual features with autoregressive training of large language models to achieve multimodal understanding.
Graph Diffusion Transformers are In-Context Molecular Designers
Gang Liu (University of Notre Dame), Meng Jiang (University of Notre Dame)
Drug DiscoveryTransformerDiffusion modelScore-based ModelBiomedical Data
🎯 What it does: Studied a graph diffusion Transformer (DemoDiff) conditioned on examples for unsupervised in-context learning in molecular inverse design;
Graph homophily booster: Reimagining the role of discrete features in heterophilic graph learning
Ruizhong Qiu (University of Illinois Urbana-Champaign), Hanghang Tong (University of Illinois Urbana-Champaign)
ClassificationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes the GRAPHITE framework, which introduces feature nodes corresponding to discrete features, connecting nodes with similar features indirectly through feature edges, thereby significantly enhancing graph homophily and improving node classification performance on heterogeneous graphs without altering the model architecture.
Graph Random Features for Scalable Gaussian Processes
Matthew Zhang (University of Cambridge), Isaac Reid (University of Cambridge)
OptimizationComputational EfficiencyGraph
🎯 What it does: Developed a framework that leverages Graph Random Features (GRF) to enable scalable Gaussian processes for Bayesian inference and optimization on large-scale graph nodes.
Graph Representational Learning: When Does More Expressivity Hurt Generalization?
Sohir Maskey (Aleph Alpha Research), Johannes F. Lutzeyer (LIX, CNRS, cole Polytechnique, Institut Polytechnique de Paris)
Representation LearningDrug DiscoveryGraph Neural NetworkGraphBiomedical Data
🎯 What it does: This paper investigates the relationship between the expressiveness and generalization performance of graph neural networks (GNNs), proposes a pseudo-metric ζ-TMD based on graph invariants, derives a generalization bound incorporating structural similarity terms, and experimentally verifies that when expressiveness aligns with task structure, performance improves, while excessive expressiveness leads to overfitting.
Graph Signal Processing Meets Mamba2: Adaptive Filter Bank via Delta Modulation
Yehjin Shin (KAIST), Noseong Park (KAIST)
GenerationComputational EfficiencyLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose the HADES architecture, reinterpreting Mamba2 as a graph signal processing filter bank, and introducing shared filters and expert filters to achieve hierarchical adaptive filtering.
Graph Tokenization for Bridging Graphs and Transformers
Zeyuan Guo (Beijing University of Posts and Telecom), Chuan Shi (Beijing University of Posts and Telecom)
Representation LearningDrug DiscoveryGraph Neural NetworkTransformerGraph
🎯 What it does: Propose a graph Tokenization framework that converts graph structures into discrete symbolic sequences by combining reversible and deterministic structure-guided serialization (frequency-guided Euler circuit/CPP) with Byte-Pair Encoding (BPE), enabling standard Transformers to directly process graph data.
Graph-based Nearest Neighbors with Dynamic Updates via Random Walks
Nina Mishra (Amazon), Lichen Zhang (MIT)
RetrievalComputational EfficiencyGraph Neural NetworkPoint CloudGraph
🎯 What it does: Proposed a new HNSW deletion algorithm called SPatch based on random walks, which can efficiently delete points on large-scale datasets while maintaining the recall rate of approximate nearest neighbor retrieval.
Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration
Sukwon Yun (University Of North Carolina Chapel Hill), Tianlong Chen (University Of North Carolina Chapel Hill)
Graph Neural NetworkTransformerLarge Language ModelAgentic AIText
🎯 What it does: This paper proposes Graph-of-Agents (GoA), a multi-agent LLM collaboration framework based on graph structures, which selects the most relevant agents through node sampling, constructs a weighted directed graph via edge sampling, performs bidirectional message passing, and aggregates the final answer through graph pooling.
Graph-Theoretic Intrinsic Reward: Guiding RL with Effective Resistance
Jatin Chauhan (Fujitsu Research), Manohar Kaul (Fujitsu Research)
Graph Neural NetworkReinforcement LearningGraphBenchmark
🎯 What it does: Propose an effective resistance (Effective Resistance) based on spectral graph theory as an intrinsic reward in sparse reward environments, using PPO for end-to-end learning.
GraphOmni: A Comprehensive and Extensible Benchmark Framework for Large Language Models on Graph-theoretic Tasks
Hao Xu (Chinese University of Hong Kong), Tianshu Yu (Chinese University of Hong Kong)
Large Language ModelReinforcement LearningPrompt EngineeringTextGraphBenchmarkChain-of-Thought
🎯 What it does: This paper constructs a multidimensional graph theory task benchmark named GRAPHOMNI, designed to systematically evaluate the reasoning capabilities of large language models across different graph types, serialization formats, and prompting schemes.
Graphon Cross-Validation: Assessing Models on Network Data
Huimin Cheng (Boston University), Wenxuan Zhong (University of Georgia)
Computational EfficiencyHyperparameter SearchGraph Neural NetworkGraph
🎯 What it does: Propose a Graphon Cross-Validation with Random Imputation method based on random imputation for hyperparameter tuning and optimal model selection in network data.
GraphPlanner: Graph Memory-Augmented Agentic Routing for Multi-Agent LLMs
Tao Feng (University of Illinois Urbana Champaign), Jiaxuan You (University of Illinois Urbana Champaign)
OptimizationComputational EfficiencyGraph Neural NetworkLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: This paper studies the routing problem in multi-agent large language models, proposing a heterogeneous graph memory-enhanced agent router called GraphPlanner, which can automatically generate multi-step workflows for each query and execute LLMs with different roles;
GraphShield: Graph-Theoretic Modeling of Network-Level Dynamics for Robust Jailbreak Detection
Sunghee Dong (Electronics and Telecommunication Research Institute), Seongyeop Kim (Electronics and Telecommunication Research Institute)
Anomaly DetectionSafty and PrivacyAdversarial AttackGraph Neural NetworkTransformerLarge Language ModelTextBenchmark
🎯 What it does: Developed a graph theory-based internal information routing analysis method called GraphShield for detecting jailbreak attacks in large language models.
GraphUniverse: Synthetic Graph Generation for Evaluating Inductive Generalization
Louis Van Langendonck (Polytechnic University of Catalonia), Pere Barlet-Ros (Polytechnic University of Catalonia)
Data SynthesisGraph Neural NetworkTransformerGraphBenchmark
🎯 What it does: This paper proposes and implements GraphUniverse, a framework capable of generating a family of graphs with semantic consistency and controllable structural properties, for systematically evaluating the inductive generalization of graph learning models on unseen graphs;
Grasp Any Region: Towards Precise, Contextual Pixel Understanding for Multimodal LLMs
Haochen Wang (Chinese Academy of Sciences), Zhaoxiang Zhang (Chinese Academy of Sciences)
SegmentationTransformerLarge Language ModelVision Language ModelImageVideoTextMultimodalityBenchmark
🎯 What it does: Proposes the GAR model, which can precisely understand arbitrary mask regions in a single image and perform multi-region interaction reasoning.
Greater than the Sum of Its Parts: Building Substructure into Protein Encoding Models
Robert Calef (MIT), Marinka Zitnik (Harvard University)
Protein Structure PredictionSupervised Fine-TuningBiomedical DataBenchmark
🎯 What it does: Constructed the Magneton environment, including a large-scale protein substructure dataset, training framework, and evaluation benchmark, and proposed a substructure-tuning method that uses annotated substructure information to perform supervised fine-tuning on pre-trained protein encoders.
GRL-SNAM: Geometric Reinforcement Learning with Differential Hamiltonians for Navigation and Mapping in Unknown Environments
Aditya Sai Ellendula (University of Texas at Austin), Chandrajit L. Bajaj (University of Texas at Austin)
Autonomous DrivingOptimizationReinforcement LearningSimultaneous Localization and Mapping
🎯 What it does: Propose the GRL-SNAM framework, treating simultaneous navigation and mapping as Hamiltonian dynamics under local perception, using the gradient of the energy field to directly generate control actions;
GRO-RAG: Gradient-aware Re-rank Optimization for Multi-source Retrieval-Augmented Generation
Siyuan Chen (University Of Bristol), Jiechao Gao (Stanford University)
RetrievalOptimizationTransformerTextRetrieval-Augmented Generation
🎯 What it does: Proposes a training-agnostic, gradient-aware multi-source retrieval-augmented generation framework called GRO-RAG, which first selects information sources through subset optimization, then uses a single backward pass to estimate the contribution of each document to the generation loss and performs Top-k re-ranking.
Grokking in LLM Pretraining? Monitor Memorization-to-Generalization without Test
Ziyue Li (University of Maryland), Tianyi Zhou (MBZUAI)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: This paper empirically observes the local asynchronous grokking phenomenon for the first time during the pretraining of a 7B parameter Mixture-of-Experts LLM (OLMoE) at actual scale, and proposes two cost-free monitoring metrics based on expert paths that can track the model's transition from memorization to generalization without requiring instruction fine-tuning or external validation set evaluation.
Ground Slow, Move Fast: A Dual-System Foundation Model for Generalizable Vision-Language Navigation
Meng Wei (Shanghai AI Laboratory), Xihui Liu (University of Hong Kong)
Autonomous DrivingTransformerPrompt EngineeringVision Language ModelDiffusion modelImageMultimodalityBenchmark
🎯 What it does: Propose the DualVLN dual-system foundation model, separating high-level semantic reasoning from low-level action execution. System 2 uses a large-scale VLM for pixel-level target prediction and adaptive perspective adjustment, while System 1 employs a lightweight diffusion Transformer to generate continuous smooth trajectories based on implicit latent goals and high-frequency RGB inputs, achieving efficient real-time control and dynamic obstacle avoidance.
Grounding and Enhancing Informativeness and Utility in Dataset Distillation
Shaobo Wang (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
Computational EfficiencyKnowledge DistillationData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: Propose a theoretical framework defining 'informativeness' and 'utility' in dataset distillation, and build the InfoUtil method by combining maximization of informativeness and utility to generate high-quality distilled datasets.
Grounding Computer Use Agents on Human Demonstrations
Aarash Feizi (Mila - Quebec AI Institute), Sai Rajeswar (Mila - Quebec AI Institute)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelImageTextMultimodality
🎯 What it does: Constructed the largest desktop UI localization dataset, GROUNDCUA, and trained a series of GROUNDNEXT models based on Qwen2.5-VL to accomplish the precise localization task from natural language instructions to screen elements.
Grounding Generative Planners in Verifiable Logic: A Hybrid Architecture for Trustworthy Embodied AI
Feiyu Wu (Xidian University), HUI LI
Safty and PrivacyRobotic IntelligenceTransformerLarge Language ModelVision Language ModelTextMultimodalityBenchmark
🎯 What it does: Proposed the Verifiable Iterative Refinement Framework (VIRF), a neuro-symbolic architecture that couples a large language model (LM) planner with a Logic Tutor based on OWL 2 logic, for achieving safe and verifiable robot planning in home environments.
Grounding or Guessing? Visual Signals for Detecting Hallucinations in Sign Language Translation
Yasser HAMIDULLAH, Cristina España-Bonet
Anomaly DetectionExplainability and InterpretabilityTransformerVision Language ModelContrastive LearningVideoText
🎯 What it does: Proposed a visual-support-based reliability metric to detect hallucinations in sign language translation (SLT), combined with traditional text uncertainty metrics (confidence, entropy, perplexity) to enhance hallucination detection performance.
Grounding-IQA: Grounding Multimodal Language Model for Image Quality Assessment
Zheng Chen (Shanghai Jiao Tong University), Yulun Zhang (Shanghai Jiao Tong University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposes a new task paradigm called 'Grounding-IQA', enabling fine-grained image quality assessment in multimodal language models, including detailed description with key region localization (GIQA-DES) and region-based quality visual question answering (GIQA-VQA).
Group Critical-token Policy Optimization for Autoregressive Image Generation
Guohui Zhang (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)
GenerationOptimizationReinforcement Learning from Human FeedbackTransformerImageMultimodalityBenchmark
🎯 What it does: Developed a Group Critical-token Policy Optimization (GCPO) method that optimizes only critical tokens in autoregressive image generation to improve RLVR training effectiveness.
Group Representational Position Encoding
Yifan Zhang (Princeton University), Andrew C Yao
Representation LearningTransformerLarge Language ModelText
🎯 What it does: Proposed a unified position encoding framework based on group actions, GRAPE, which includes two major categories: multiplicative (SO(d) rotations) and additive (GL near-identity matrices).
Group Verification-based Policy Optimization for Interactive Coding Agents
Silong Dai (East China Normal University), Xuelong Li (China Telecom)
AI Code AssistantTransformerLarge Language ModelReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Proposed a new reinforcement learning algorithm called GVPO, which combines final outcome rewards with intermediate execution feedback for training interactive coding agents;
Group-Normalized Implicit Value Optimization for Language Models
Yunseon Choi (KAIST AI), Kee-Eung Kim (KAIST AI)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose GN-IVO, a RL fine-tuning method without critic, which implicitly learns step-level value through group normalization distribution matching, addressing the sparse reward credit assignment problem in LLM generation.
Group-Relative REINFORCE Is Secretly an Off-Policy Algorithm: Demystifying Some Myths About GRPO and Its Friends
Chaorui Yao (University of California, Los Angeles), Bolin Ding (Alibaba Group)
Reinforcement LearningText
🎯 What it does: This paper provides a theoretical analysis of Group-Relative REINFORCE, revealing for the first time that its essence is an off-policy algorithm. Based on this, it proposes two general improvement principles: regularization updates and data distribution adjustment. Furthermore, it unifies and reinterprets algorithms such as GRPO, OPMD, and AsymRE, and introduces data-weighting strategies such as RED-DROP and RED-WEIGHT.
Grouping Nodes with known Value Differences: A lossless UCT-based Abstraction Algorithm
Robin Schmöcker (Leibniz University Hannover), Bodo Rosenhahn (Leibniz University Hannover)
OptimizationReinforcement Learning
🎯 What it does: Propose the KVDA-UCT algorithm, extending OGA-UCT by introducing known value difference abstraction to improve the sample efficiency of MCTS.
GT-Space: Enhancing Heterogeneous Collaborative Perception with Ground Truth Feature Space
Wentao Wang (Sun Yat-sen University), Guang Tan (Sun Yat-sen University)
Object DetectionAutonomous DrivingTransformerContrastive LearningImageMultimodalityPoint Cloud
🎯 What it does: Proposes the GT-Space framework, achieving heterogeneous collaborative perception across different sensing modalities; by constructing a shared feature space based on real labels, simplifying the feature alignment and fusion process.
GTA1: GUI Test-time Scaling Agent
Yan Yang (Salesforce Ai Research), Junnan Li (Salesforce Ai Research)
Data-Centric LearningRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIMultimodality
🎯 What it does: Propose the GTA1 framework, which integrates planning with test-time expansion and a GUI grounding model that directly locates coordinates using RL, to construct a two-stage GUI agent.
GTM: A General Time-series Model for Enhanced Representation Learning of Time-Series data
Cheng HE, Patrick Lee
Anomaly DetectionRepresentation LearningTransformerTime Series
🎯 What it does: Proposed a general-purpose time series model GTM that can handle various generation tasks without task-specific modifications.
GTool: Graph Enhanced Tool Planning with Large Language Model
Wenjie Chen (Chinese Academy of Sciences), Jingping Bi (Chinese Academy of Sciences)
Graph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextGraph
🎯 What it does: Propose GTool, which realizes tool planning by constructing request-specific tool graphs and integrating graph neural networks with large language models.
GTR-Bench: Evaluating Geo-Temporal Reasoning in Vision-Language Models
Qinghongbing Xie (Tsinghua University), Long ZENG
Large Language ModelVision Language ModelImageVideoMultimodalityBenchmark
🎯 What it does: Proposed and implemented the GTR-Bench benchmark to evaluate the ability of vision-language models to perform geospatial-temporal reasoning by integrating maps and videos in large-scale multi-camera networks.
Guaranteed Simply Connected Mesh Reconstruction from an Unorganized Point Cloud
Liyan Chen (University of Texas at Austin), Qixing Huang (University of Texas at Austin)
GenerationOptimizationPoint CloudMeshComputed Tomography
🎯 What it does: Reconstruct a closed, simply connected 3D mesh from unordered noisy point clouds while ensuring topological correctness.
GuardAlign: Test-time Safety Alignment in Multimodal Large Language Models
Xingyu Zhu (University of Science and Technology of China), Xiangnan He (Tianjin University)
Safty and PrivacyTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Developed GuardAlign, a training-agnostic input detection and decoding calibration framework to enhance the safety of large vision-language models.
GUI-Shift: Enhancing VLM-Based GUI Agents through Self-supervised Reinforcement Learning
Longxi Gao (Beijing University of Posts and Telecommunications), Mengwei Xu (Beijing University of Posts and Telecommunications)
TransformerSupervised Fine-TuningReinforcement LearningImage
🎯 What it does: Proposed the K-step GUI Transition self-supervised inverse dynamics task and developed the GUI-Shift RL framework to train VLM agents using unlabeled GUI trajectories.
Guidance Matters: Rethinking the Evaluation Pitfall for Text-to-Image Generation
Dian Xie (Hong Kong University of Science and Technology (Guangzhou)), Zeke Xie (Hong Kong University of Science and Technology (Guangzhou))
GenerationDiffusion modelImageMultimodality
🎯 What it does: This paper explores and reveals the 'guidance scale bias' problem in text-to-image generation evaluation, proposes the GA-Eval framework based on effective guidance scale, re-evaluates existing diffusion guidance methods using this framework, and further designs the Transcendent Diffusion Guidance (TDG) method that is easily misjudged.
Guidance Watermarking for Diffusion Models
Enoal Gesny (Inria), Vivien Chappelier (LABEL4.AI)
GenerationDiffusion modelImage
🎯 What it does: This paper proposes a gradient-guided diffusion model watermark embedding method that converts any post-watermark into an embedded component during the generation process without requiring retraining of the model.
GUIDE: Gated Uncertainty-Informed Disentangled Experts for Long-tailed Recognition
Yuan Dong (University of Science and Technology of China), Pengkun Wang (University of Science and Technology of China)
RecognitionMixture of ExpertsImage
🎯 What it does: Propose the GUIDE framework, addressing the multi-expert architecture bottleneck in long-tailed recognition through hierarchical decoupling;
Guided Flow Policy: Learning from High-Value Actions in Offline Reinforcement Learning
Franki NGUIMATSIA TIOFACK, Justin Carpentier (Inria and DI-ENS, PSL Research University Sorbonne Universite, CNRS, ISIR)
Reinforcement LearningFlow-based ModelImageTabularBenchmark
🎯 What it does: Developed an offline reinforcement learning algorithm called Guided Flow Policy, combining multi-step flow matching strategy with a single-step actor, leveraging value-aware behavior cloning to enhance utilization of high-value actions in the dataset.
Guided Policy Optimization under Partial Observability
Yueheng Li (Peking University), Zongqing Lu (Peking University)
OptimizationReinforcement Learning
🎯 What it does: Propose the Guided Policy Optimization (GPO) framework, which jointly trains the guide and learner to utilize privileged information in partially observable environments.
Guided Query Refinement: Multimodal Hybrid Retrieval with Test-Time Optimization
Omri Uzan (Stanford University), Ariel Gera (IBM Research)
RetrievalOptimizationComputational EfficiencyTransformerMultimodalityBenchmark
🎯 What it does: Propose a hybrid retrieval method GQR based on test-time optimization, which refines the query vectors of a visual retriever using the scores from a lightweight text retriever.
Guided Speculative Inference for Efficient Test-Time Alignment of LLMs
Jonathan Geuter (Harvard SEAS Kempner Institute), David Alvarez-Melis (Harvard SEAS Kempner Institute)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose the Guided Speculative Inference (GSI) method, which performs reward-guided best-N sampling during inference for large language models and combines speculative decoding for acceleration.
GuidedBench: Measuring and Mitigating the Evaluation Discrepancies of In-the-wild LLM Jailbreak Methods
Ruixuan Huang (Hong Kong University of Science and Technology), Shuai Wang (Hong Kong University of Science and Technology)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose GuidedBench, which includes a fine-grained dataset for LLM malicious question answering and a case-guided evaluation system called GuidedEval, to assess the effectiveness of jailbreak methods.
GuidedSampling: Steering LLMs Towards Diverse Candidate Solutions at Inference-Time
Divij Handa (Arizona State University), Chitta Baral (Arizona State University)
GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Propose a new algorithm called GUIDEDSAMPLING, which first generates diverse concepts or theorems during the exploration phase, and then uses these concepts to generate candidate solutions during the generation phase, thereby improving the diversity and accuracy of solutions.
GuirlVG: Incentivize GUI Visual Grounding via Empirical Exploration on Reinforcement Learning
Weitai Kang (University of Illinois Chicago), Yan Yan (University of Illinois Chicago)
TransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Propose GuirlVG, a GUI visual localization reinforcement learning method based on GRPO, which significantly outperforms SFT using only 5.2K samples through systematic experiments and new technologies.
Gumbel Distillation for Parallel Text Generation
Chi Zhang (University of Texas at Austin), qiang liu
GenerationComputational EfficiencyKnowledge DistillationTransformerDiffusion modelText
🎯 What it does: Introduces the Gumbel Distillation method, mapping the sampling process of the autoregressive teacher model into supervised learning noise-text pairs via the Gumbel-Max trick, enabling the parallel decoder to learn the joint distribution.
h-MINT: Modeling Pocket-Ligand Binding with Hierarchical Molecular Interaction Network
Yanru Qu (University of Illinois Urbana-Champaign), Ge Liu (University of Illinois Urbana-Champaign)
Drug DiscoveryGraph Neural NetworkTransformerBiomedical Data
🎯 What it does: Developed a protein-ligand interaction modeling framework based on overlapping BPE tokenization and hierarchical molecular interaction networks (h-MINT), significantly improving binding affinity prediction and virtual screening performance.
H$^3$DP: Triply‑Hierarchical Diffusion Policy for Visuomotor Learning
Yiyang Lu (Institute for Interdisciplinary Information Sciences, Tsinghua University), Huazhe Xu (Institute for Interdisciplinary Information Sciences, Tsinghua University)
Depth EstimationRobotic IntelligenceConvolutional Neural NetworkDiffusion modelAuto EncoderMultimodality
🎯 What it does: Propose a three-level visual motion strategy framework H³DP, specifically designed to address the coupling problem between visual perception and action generation.
H2OFlow: Grounding Human-Object Affordances with 3D Generative Models and Dense Diffused Flows
Harry Zhang (MIT), Luca Carlone (MIT)
GenerationData SynthesisDiffusion modelFlow-based ModelPoint Cloud
🎯 What it does: Utilizing synthetic 3D human-robot interaction data, learning complete affinities (contact, orientation, spatial) in 3D human-robot interaction through point cloud dense diffused flows.
HackWorld: Evaluating Computer-Use Agents on Exploiting Web Application Vulnerabilities
Xiaoxue Ren (Zhejiang University), Terry Yue Zhuo (Monash University)
Safty and PrivacyLarge Language ModelAgentic AITextMultimodalityBenchmark
🎯 What it does: Proposed the HackWorld framework, systematically evaluated the ability of computer-using agents (CUA) to identify and exploit real Web application vulnerabilities through visual interaction, and conducted experiments in 36 CTF environments containing real vulnerabilities.
Half-order Fine-Tuning for Diffusion Model: A Recursive Likelihood Ratio Optimizer
Tao Ren (Peking University), Yijie Peng (Peking University)
GenerationReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningPrompt EngineeringDiffusion modelImageVideoTextChain-of-Thought
🎯 What it does: Proposed a half-order gradient estimator called Recursive Likelihood Ratio (RLR) Optimizer for efficiently fine-tuning diffusion models, combined with a multi-scale prompting technique named Diffusive Chain-of-Thought (DCoT).
Hallucination Begins Where Saliency Drops
Xiaofeng Zhang (Shanghai Jiaotong University), Hao Tang (Peking University)
Explainability and InterpretabilityComputational EfficiencyLarge Language ModelReinforcement LearningVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose a LVLMs-Saliency metric based on gradient attention, and integrate Saliency-Guided Rejection Sampling with Local Coherence Reinforcement mechanisms during inference, significantly reducing hallucinations in image question answering and description generation.
Hallucination Reduction with CASAL: Contrastive Activation Steering for Amortized Learning
Wannan Yang (Meta Superintelligence Labs), Diego Garcia-Olano (Meta Superintelligence Labs)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningTextMultimodality
🎯 What it does: This paper proposes a 'CASAL' training framework based on contrastive activation modulation, which directly embeds the model's knowledge boundary into weights by using a representation layer loss on a single-layer network, enabling LLMs to self-denial when encountering unknown questions and reducing hallucinations.
Hallucination-aware Intermediate Representation Edit in Large Vision-Language Models
Wei Suo (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)
Explainability and InterpretabilityRepresentation LearningVision Language ModelContrastive LearningImageText
🎯 What it does: Propose the HIRE framework, which detects and edits intermediate representations of large vision-language models to eliminate or regulate hallucinations without requiring model retraining or dual reasoning.
HalluGuard: Demystifying Data-Driven and Reasoning-Driven Hallucinations in LLMs
Xinyue Zeng (Virginia Tech), Dawei Zhou (Virginia Tech)
Explainability and InterpretabilityTransformerTextBenchmark
🎯 What it does: Proposes a unified theoretical framework for Hallucination Risk Bound, and introduces the HALLUGUARD scoring model based on NTK to simultaneously detect data-driven and reasoning-driven hallucinations, followed by extensive validation on 10 benchmarks.
HAMLET: A Hierarchical and Adaptive Multi-Agent Framework for Live Embodied Theatrics
Shufan Jiang (East China University of Science and Technology), Xuelong Li (Institute of Artificial Intelligence TeleAI China Telecom)
TransformerLarge Language ModelAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Proposed a multi-agent framework named HAMLET for automatically generating script blueprints and enabling autonomous decision-making, emotional expression, and interaction with the physical environment by AI actors in real-time performances.
HAMLET: Switch Your Vision-Language-Action Model into a History-Aware Policy
Myungkyu Koo (KAIST), Jinwoo Shin (KAIST)
Robotic IntelligenceTransformerVision-Language-Action ModelContrastive LearningMultimodality
🎯 What it does: By incorporating learnable temporal tokens and a lightweight memory module into a pre-trained vision-language-action model, the model can utilize historical context for action prediction.
HardcoreLogic: Challenging Large Reasoning Models with Long-tail Logic Puzzle Games
Jingcong Liang (Fudan University), zhongyu wei
Large Language ModelTextBenchmark
🎯 What it does: Proposed a logic puzzle benchmark named HardcoreLogic, containing over 5,000 puzzles generated by three-dimensional long-tail transformations (IC, UE, UP) across multiple game genres.
Harder Is Better: Boosting Mathematical Reasoning via Difficulty-Aware GRPO and Multi-Aspect Question Reformulation
Yanqi Dai (Renmin University of China), Zhiwu Lu (Renmin University of China)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose the MathForge framework, combining difficulty-aware Group Policy Optimization (DGPO) and multi-dimensional problem rewriting (MQR) to enhance the performance of large language models on mathematical reasoning tasks.
HARDTESTGEN: A High-Quality RL Verifier Generation Pipeline for LLM Algorithimic Coding
Zhongmou He (Carnegie Mellon University), Lei Li
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Propose an LLM-based test generation pipeline called HARDTESTGEN, and construct a high-quality test set named HARDTESTS with 26.6k problems for reinforcement learning and rejection sampling in post-training of LLMs.
Harmonized Cone for Feasible and Non-conflict Directions in Training Physics-Informed Neural Networks
Dohyun Bu (Korea Advanced Institute of Science and Technology), Jong-Seok Lee (Korea Advanced Institute of Science and Technology)
OptimizationBenchmarkPhysics Related
🎯 What it does: Introduces the concept of harmonized cone to ensure gradient directions in PINN training are both feasible and conflict-free, and designs the HARMONIC algorithm based on this concept.
HarmonyGNNs: Harmonizing Heterophily and Homophily in GNNs via Self-Supervised Node Encoding
Rui Xue (North Carolina State University), Tianfu Wu (North Carolina State University)
Computational EfficiencyRepresentation LearningGraph Neural NetworkTransformerContrastive LearningGraph
🎯 What it does: Propose the HarmonyGNNs framework, integrating teacher-student prediction with adaptive node masking to unify self-supervised learning for heterogeneous and homogeneous graphs.
Harnessing Hyperbolic Geometry for Harmful Prompt Detection and Sanitization
Igor Maljkovic (University of Genoa), Fabio Roli (University of Genoa)
Anomaly DetectionExplainability and InterpretabilityLarge Language ModelText
🎯 What it does: This paper proposes two modules: HyPE for detecting harmful prompts, and HyPS for interpreting and purifying detected harmful prompts, thereby protecting vision-language models.
Harnessing Temporal Databases for Systematic Evaluation of Factual Time-Sensitive Question-Answering in LLMs
Soyeon Kim (KAIST), Steven Euijong Whang (William & Mary)
Explainability and InterpretabilityTransformerLarge Language ModelTabularTime SeriesBiomedical DataBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed TDBench, a benchmark framework that automatically constructs time-sensitive question-answer (TSQA) pairs using a time series database, and introduced the time accuracy evaluation metric;
HARP: Hallucination Detection via Reasoning Subspace Projection
Junjie Hu (Huazhong University of Science and Technology), Dongbo Shan (Huazhong University of Science and Technology)
Anomaly DetectionTransformerLarge Language ModelText
🎯 What it does: Propose the HARP framework, which detects hallucinations by projecting the reasoning subspace from the hidden states of LLMs.
Harpoon: Generalised Manifold Guidance for Conditional Tabular Diffusion
Aditya Shankar (Delft University of Technology), Lydia Y. Chen
GenerationData SynthesisDiffusion modelTabular
🎯 What it does: Proposed a conditional tabular diffusion model named HARPOON, achieving multiple constraint generation during training-free scenarios through manifold guidance.
HATSolver: Learning Gröbner Bases with Hierarchical Attention Transformers
Mohamed Malhou (FAIR, Meta Superintelligence Labs), Kristin E. Lauter (FAIR, Meta Superintelligence Labs)
Computational EfficiencyData-Centric LearningTransformer
🎯 What it does: Proposed the Hierarchical Attention Transformer (HAT) model for computing Gröbner bases of multivariate polynomial systems;
HBO: Hierarchical Balancing Optimization for Fine-Tuning Large Language Models
Weixuan Wang (University of Edinburgh), Alexandra Birch (University of Edinburgh)
OptimizationSupervised Fine-TuningReinforcement LearningText
🎯 What it does: To address the issues of imbalance and heterogeneity across datasets and within individual datasets during the fine-tuning of large language models, the Hierarchical Balancing Optimization (HBO) framework is proposed, enabling the model to autonomously adjust global and local data sampling ratios during training.
HDR-NSFF: High Dynamic Range Neural Scene Flow Fields
Shin Dong-Yeon (KAIST), Tae-Hyun Oh (KAIST)
RestorationNeural Radiance FieldOptical FlowVideo
🎯 What it does: Propose the HDR-NSFF framework, which reconstructs dynamic HDR radiance fields from alternating exposure monocular videos using a 4D spatiotemporal neural field.
Heads collapse, features stay: Why Replay needs big buffers
Giulia Lanzillotta (ETH Zurich), Thomas Hofmann (ETH Zurich)
ClassificationRepresentation LearningMeta LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: Studied the differences between deep forgetting (feature space separability) and shallow forgetting (classifier alignment) in continuous learning, and explained the impact of experience replay buffer size on both types of forgetting through the neural collapse framework.
Healthcare Insurance Fraud Detection via Continual Fiedler Vector Graph Model
Yehan Zhang (South China University of Technology), Shengfeng He (Singapore Management University)
Anomaly DetectionGraph Neural NetworkAuto EncoderGraphFinance Related
🎯 What it does: Designed a continuous learning graph model named ConFVG for real-time identification in medical insurance fraud detection under low-labeling and non-stationary environments, combining a Fiedler vector-guided graph autoencoder with a subgraph attention fusion module.
HEAPr: Hessian-based Efficient Atomic Expert Pruning in Output Space
Ke Li (Zhejiang University), Wenxiao Wang (Zhejiang University)
Computational EfficiencyKnowledge DistillationMixture of ExpertsText
🎯 What it does: Propose a second-order information-based MoE atomic expert pruning method called HEAPr, which first splits experts into indivisible atomic experts, then estimates the importance of each atomic expert in the output space via the Fisher information matrix, achieving pruning with only two forward and one backward pass;
Hedonic Neurons: A Mechanistic Mapping of Latent Coalitions in Transformer MLPs
Tanya Chowdhury (University of Massachusetts Amherst), James Allan (University of Massachusetts Amherst)
Explainability and InterpretabilityTransformerSupervised Fine-TuningText
🎯 What it does: Propose a mechanism explanation framework based on game theory, treating neurons in the Transformer MLP layer as players, leveraging synergy effects to discover and track stable neuron coalitions, revealing collaborative computational units within the model.
HEEGNet: Hyperbolic Embeddings for EEG
Shanglin Li (Advanced Telecommunications Research Institute International), Ziheng Chen (University of Trento)
ClassificationRecognitionDomain AdaptationBiomedical Data
🎯 What it does: Developed HEEGNet, a hybrid hyperbolic network combining Euclidean and hyperbolic encoders, and introduced a two-stage domain adaptation (DSMDBN) to achieve robust improvements in cross-domain EEG decoding.
HEIST: A Graph Foundation Model for Spatial Transcriptomics and Proteomics Data
Hiren Madhu (Yale University), Rex Ying (Yale University)
Graph Neural NetworkTransformerAuto EncoderContrastive LearningGraphBiomedical DataAlzheimer's Disease
🎯 What it does: Proposed and implemented a hierarchical graph Transformer model named HEIST for spatial transcriptomics and proteomics data, capable of simultaneously modeling spatial adjacency between cells and gene co-expression networks within cells, and generating transferable cell and gene embeddings through cross-layer message passing.
Helix: Evolutionary Reinforcement Learning for Open-Ended Scientific Problem Solving
Chang Su (Bosch (China) Investment Co Ltd), Jun Zhu (Tsinghua University)
OptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTabularBenchmark
🎯 What it does: Developed a hierarchical evolutionary reinforcement learning framework called HELIX, leveraging LLMs to achieve iterative optimization in open-ended scientific problems through experience learning, population diversity maintenance, and contextual prompting.
Helmsman: Autonomous Synthesis of Federated Learning Systems via Collaborative LLM Agents
Haoyuan Li (Eindhoven University of Technology), Aaqib Saeed (Eindhoven University of Technology)
Federated LearningLarge Language ModelAgentic AIBenchmarkRetrieval-Augmented Generation
🎯 What it does: Developed a multi-agent system called Helmsman, which can automatically complete planning, modular coding, and closed-loop evaluation based on users' high-level federated learning (FL) requirements, ultimately generating a directly deployable federated learning system.
Hessian-Enhanced Token Attribution (HETA): Interpreting Autoregressive LLMs
Vishal Pramanik (University of Florida), Sumit Kumar Jha (University of Florida)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose an explanation method named HETA, which provides word-level attribution for each generated word in autoregressive language models with only decoder structure, based on semantic flow, Hessian second-order sensitivity, and KL information gain.
Heterogeneous Agent Q-weighted Policy Optimization
Bor-Jiun Lin (National Taiwan University), Chun-Yi Lee (National Taiwan University)
OptimizationReinforcement LearningDiffusion modelBenchmark
🎯 What it does: Proposes the Heterogeneous-Agent Q-weighted Optimization (HAQO) framework, combining serialized advantage updates, Q-weighted variational objectives, and entropy regularization to address the stability and expressiveness trade-off in heterogeneous multi-agent reinforcement learning.
Heterogeneous Federated Fine-Tuning with Parallel One-Rank Adaptation
Zikai Zhang (University of Nevada Reno), Jiahao Xu (University of Nevada Reno)
Federated LearningComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextFinance Related
🎯 What it does: Propose a federated fine-tuning framework called Fed-PLoRA suitable for heterogeneous resource environments, achieving zero initialization noise and low aggregation noise by utilizing parallel first-order low-rank modules (PLoRA) and the Select-N-Fold strategy.
HeurekaBench: A Benchmarking Framework for AI Co-scientist
Siba Smarak Panigrahi (École Polytechnique Fédérale de Lausanne), Maria Brbic (ETH Zurich)
AI Code AssistantTransformerLarge Language ModelTextBiomedical DataBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the HEUREKABENCH framework for constructing an AI co-scientist benchmark, which uses multi-LLM pipelines to extract reproducible scientific insights from research papers and code repositories, generate open-ended research questions, and further evaluate AI agents in the field of single-cell biology.
HeuriGym: An Agentic Benchmark for LLM-Crafted Heuristics in Combinatorial Optimization
Hongzheng Chen (Cornell University), Zhiru Zhang (Cornell University)
OptimizationAgentic AIPrompt EngineeringBenchmark
🎯 What it does: Proposes HeuriGym, an evaluation framework for LLMs, to assess their ability to generate and iteratively improve heuristic algorithms for combinatorial optimization problems.
Hey, That's My Model! Introducing Chain & Hash, An LLM Fingerprinting Technique
Mark Russinovich (Microsoft), Ahmed Salem (Microsoft)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the Chain & Hash framework, utilizing chain hashing technology to embed and detect verifiable and black-box identifiable fingerprints in LLMs.
HFSTI-Net: Hierarchical Frequency-spatial-temporal Interactions for Video Polyp Segmentation
Yuanqin He (Shenzhen University), Jing Qin (Hong Kong Polytechnic University)
SegmentationConvolutional Neural NetworkTransformerVideoBiomedical Data
🎯 What it does: Proposed a new network called HFSTI-Net for video polyp segmentation, specifically addressing the two major challenges of shape collapse and event forgetting.
HGNet: Scalable Foundation Model for Automated Knowledge Graph Generation from Scientific Literature
Devvrat Joshi (Imperial College London), Islem Rekik (Imperial College London)
RecognitionGenerationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed a two-phase automated knowledge graph construction framework (Z-NERD for entity recognition, HGNet for hierarchical relation extraction), and released a large-scale multi-domain hierarchical relation extraction benchmark dataset SPHERE.
HiCache: A Plug-in Scaled-Hermite Upgrade for Taylor-Style Cache-then-Forecast Diffusion Acceleration
Liang Feng (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
GenerationComputational EfficiencyTransformerDiffusion modelImageVideoTextMultimodality
🎯 What it does: Proposed a training-free acceleration framework called HiCache, which enhances feature prediction accuracy by employing Hermite polynomials in the feature cache-prediction link, significantly accelerating diffusion model inference.
Hidden Breakthroughs in Language Model Training
Sara Kangaslahti (Harvard University), Naomi Saphra (Harvard University)
Explainability and InterpretabilityRepresentation LearningText
🎯 What it does: This paper proposes POLCA, a loss decomposition method based on a low-rank training subspace. By splitting the loss curve of individual samples along specific basis vectors during training, it identifies hidden phase transitions (concept breakthroughs) and performs unsupervised clustering, thereby revealing interpretable skills in the model's learning process.
HiddenEcho: Mitigating Noise Amplification in Differentially Private LLMs with Hidden-State Correction
Wenhao Li (South China University of Technology), Lei Yang (South China University of Technology)
Federated LearningSafty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Proposed the HiddenEcho framework, achieving noise reduction and communication compression for differential privacy large language models (LLMs) through client-side noise correction and server-side hidden layer information.
HiDrop: Hierarchical Vision Token Reduction in MLLMs via Late Injection, Concave Pyramid Pruning, and Early Exit
Hao Wu (Eastern Institute of Technology), Xiaoyu Shen (Eastern Institute of Technology)
Computational EfficiencyLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Propose the HiDrop framework, which significantly reduces visual computational load by hierarchically pruning visual tokens in multimodal large language models.